GPT-5 Pro cracks immunologist's shelved T-cell puzzle
According to OpenAI, GPT-5 Pro explained a 2022 T-cell result that immunologist Derya Unutmaz's lab had set aside, a case of a model offering a mechanistic idea that domain experts had missed.
OpenAI published a case study about Derya Unutmaz, an immunologist and professor at The Jackson Laboratory and the University of Connecticut. In late 2025, he used GPT-5 Pro to reopen an experiment his lab had shelved three years earlier.
In 2022, the team grew T cells early in their development either in low glucose or with deoxyglucose, a molecule that blocks the cell's use of glucose. They expected both conditions to give similar results. Instead, the deoxyglucose cells overwhelmingly became inflammatory cells, and the effect lasted after the molecule was removed. The low-glucose cells showed a much weaker shift. A lack of energy alone couldn't explain the gap, and the lab moved on.
After Unutmaz uploaded the data, GPT-5 Pro suggested that deoxyglucose interferes with the production of IL-2, a protein that normally stops T cells from becoming inflammatory Th17 cells. Without IL-2, more cells could take that path. Unutmaz said the idea "retrospectively, makes perfect sense." It came from just outside his own field, and no one in his lab had spotted it.
He also asked the model to simulate a separate experiment on CD8+ T cells that target lymphoma. According to Unutmaz, it correctly predicted that the cells would be better at killing lymphoma cells, even though those results had not been published.
The account comes from OpenAI and is told through Unutmaz's own words. The source does not say whether the IL-2 hypothesis has been tested in the lab or peer-reviewed. Unutmaz stressed that experts are still needed to judge whether an AI-generated idea is plausible and matters.